Data

A run reads two CSV files: dataset.csv, one row per sample with its image, its targets and any metadata, and splits.csv, which puts each sample in train, tune or a test split, per fold. Three readers take what they need from the same dataset.csv:

• soma.Cohort holds each sample’s identity, targets and folds, and no file paths. It checks the splits when it is built, and refuses a patient whose samples fall in two splits of one fold.

• soma.ImageManifest holds the image paths that extraction reads.

• soma.AnnotationManifest holds the supervision files of segmentation and detection: label masks and point files.

Features enter a run through a source. soma.extract() reads the image manifest (and the annotation manifest, to sample regions from whole slides), tiles and encodes the images, and returns the source the run trains on, keyed by sample_id. Features computed elsewhere enter as a source too; see Bring your own features.

• Dataset: the columns of both files, the split rules, and what makes two cohorts the same experiment

• Curation: functions that write these files for supported public datasets

• Preprocessing: tissue masking, tiling, spacing, and regions sampled from annotations